Improving Generalization in Meta-Learning via Meta-Gradient Augmentation

Meta-learning methods typically follow a two-loop framework, where each loop potentially suffers from notorious overfitting, hindering rapid adaptation and generalization to new tasks. Existing schemes solve it by enhancing the mutual-exclusivity or diversity of training samples, but these data manipulation strategies are data-dependent and insufficiently flexible. This work alleviates overfitting in meta-learning from the perspective of gradient regularization and proposes a data-independent \textbf{M}eta-\textbf{G}radient \textbf{Aug}mentation (\textbf{MGAug}) method. The key idea is to first break the rote memories by network pruning to address memorization overfitting in the inner loop, and then the gradients of pruned sub-networks naturally form the high-quality augmentation of the meta-gradient to alleviate learner overfitting in the outer loop. Specifically, we explore three pruning strategies, including \textit{random width pruning}, \textit{random parameter pruning}, and a newly proposed \textit{catfish pruning} that measures a Meta-Memorization Carrying Amount (MMCA) score for each parameter and prunes high-score ones to break rote memories as much as possible. The proposed MGAug is theoretically guaranteed by the generalization bound from the PAC-Bayes framework. In addition, we extend a lightweight version, called MGAug-MaxUp, as a trade-off between performance gains and resource overhead. Extensive experiments on multiple few-shot learning benchmarks validate MGAug's effectiveness and significant improvement over various meta-baselines. The code is publicly available at \url{https://github.com/xxLifeLover/Meta-Gradient-Augmentation}.

MetaViewer: Towards A Unified Multi-View Representation

Existing multi-view representation learning methods typically follow a specific-to-uniform pipeline, extracting latent features from each view and then fusing or aligning them to obtain the unified object representation. However, the manually pre-specify fusion functions and view-private redundant information mixed in features potentially degrade the quality of the derived representation. To overcome them, we propose a novel bi-level-optimization-based multi-view learning framework, where the representation is learned in a uniform-to-specific manner. Specifically, we train a meta-learner, namely MetaViewer, to learn fusion and model the view-shared meta representation in outer-level optimization. Start with this meta representation, view-specific base-learners are then required to rapidly reconstruct the corresponding view in inner-level. MetaViewer eventually updates by observing reconstruction processes from uniform to specific over all views, and learns an optimal fusion scheme that separates and filters out view-private information. Extensive experimental results in downstream tasks such as classification and clustering demonstrate the effectiveness of our method.

Fine-Grained Classification with Noisy Labels

Learning with noisy labels (LNL) aims to ensure model generalization given a label-corrupted training set. In this work, we investigate a rarely studied scenario of LNL on fine-grained datasets (LNL-FG), which is more practical and challenging as large inter-class ambiguities among fine-grained classes cause more noisy labels. We empirically show that existing methods that work well for LNL fail to achieve satisfying performance for LNL-FG, arising the practical need of effective solutions for LNL-FG. To this end, we propose a novel framework called stochastic noise-tolerated supervised contrastive learning (SNSCL) that confronts label noise by encouraging distinguishable representation. Specifically, we design a noise-tolerated supervised contrastive learning loss that incorporates a weight-aware mechanism for noisy label correction and selectively updating momentum queue lists. By this mechanism, we mitigate the effects of noisy anchors and avoid inserting noisy labels into the momentum-updated queue. Besides, to avoid manually-defined augmentation strategies in contrastive learning, we propose an efficient stochastic module that samples feature embeddings from a generated distribution, which can also enhance the representation ability of deep models. SNSCL is general and compatible with prevailing robust LNL strategies to improve their performance for LNL-FG. Extensive experiments demonstrate the effectiveness of SNSCL.

Self-Filtering: A Noise-Aware Sample Selection for Label Noise with Confidence Penalization

Sample selection is an effective strategy to mitigate the effect of label noise in robust learning. Typical strategies commonly apply the small-loss criterion to identify clean samples. However, those samples lying around the decision boundary with large losses usually entangle with noisy examples, which would be discarded with this criterion, leading to the heavy degeneration of the generalization performance. In this paper, we propose a novel selection strategy, \textbf{S}elf-\textbf{F}il\textbf{t}ering (SFT), that utilizes the fluctuation of noisy examples in historical predictions to filter them, which can avoid the selection bias of the small-loss criterion for the boundary examples. Specifically, we introduce a memory bank module that stores the historical predictions of each example and dynamically updates to support the selection for the subsequent learning iteration. Besides, to reduce the accumulated error of the sample selection bias of SFT, we devise a regularization term to penalize the confident output distribution. By increasing the weight of the misclassified categories with this term, the loss function is robust to label noise in mild conditions. We conduct extensive experiments on three benchmarks with variant noise types and achieve the new state-of-the-art. Ablation studies and further analysis verify the virtue of SFT for sample selection in robust learning.

Learning to Rectify for Robust Learning with Noisy Labels

Label noise significantly degrades the generalization ability of deep models in applications. Effective strategies and approaches, \textit{e.g.} re-weighting, or loss correction, are designed to alleviate the negative impact of label noise when training a neural network. Those existing works usually rely on the pre-specified architecture and manually tuning the additional hyper-parameters. In this paper, we propose warped probabilistic inference (WarPI) to achieve adaptively rectifying the training procedure for the classification network within the meta-learning scenario. In contrast to the deterministic models, WarPI is formulated as a hierarchical probabilistic model by learning an amortization meta-network, which can resolve sample ambiguity and be therefore more robust to serious label noise. Unlike the existing approximated weighting function of directly generating weight values from losses, our meta-network is learned to estimate a rectifying vector from the input of the logits and labels, which has the capability of leveraging sufficient information lying in them. This provides an effective way to rectify the learning procedure for the classification network, demonstrating a significant improvement of the generalization ability. Besides, modeling the rectifying vector as a latent variable and learning the meta-network can be seamlessly integrated into the SGD optimization of the classification network. We evaluate WarPI on four benchmarks of robust learning with noisy labels and achieve the new state-of-the-art under variant noise types. Extensive study and analysis also demonstrate the effectiveness of our model.

Learning Transferable Parameters for Unsupervised Domain Adaptation

Unsupervised domain adaptation (UDA) enables a learning machine to adapt from a labeled source domain to an unlabeled domain under the distribution shift. Thanks to the strong representation ability of deep neural networks, recent remarkable achievements in UDA resort to learning domain-invariant features. Intuitively, the hope is that a good feature representation, together with the hypothesis learned from the source domain, can generalize well to the target domain. However, the learning processes of domain-invariant features and source hypothesis inevitably involve domain-specific information that would degrade the generalizability of UDA models on the target domain. In this paper, motivated by the lottery ticket hypothesis that only partial parameters are essential for generalization, we find that only partial parameters are essential for learning domain-invariant information and generalizing well in UDA. Such parameters are termed transferable parameters. In contrast, the other parameters tend to fit domain-specific details and often fail to generalize, which we term as untransferable parameters. Driven by this insight, we propose Transferable Parameter Learning (TransPar) to reduce the side effect brought by domain-specific information in the learning process and thus enhance the memorization of domain-invariant information. Specifically, according to the distribution discrepancy degree, we divide all parameters into transferable and untransferable ones in each training iteration. We then perform separate updates rules for the two types of parameters. Extensive experiments on image classification and regression tasks (keypoint detection) show that TransPar outperforms prior arts by non-trivial margins. Moreover, experiments demonstrate that TransPar can be integrated into the most popular deep UDA networks and be easily extended to handle any data distribution shift scenarios.

Attentional Prototype Inference for Few-Shot Semantic Segmentation

This paper aims to address few-shot semantic segmentation. While existing prototype-based methods have achieved considerable success, they suffer from uncertainty and ambiguity caused by limited labelled examples. In this work, we propose attentional prototype inference (API), a probabilistic latent variable framework for few-shot semantic segmentation. We define a global latent variable to represent the prototype of each object category, which we model as a probabilistic distribution. The probabilistic modeling of the prototype enhances the model's generalization ability by handling the inherent uncertainty caused by limited data and intra-class variations of objects. To further enhance the model, we introduce a local latent variable to represent the attention map of each query image, which enables the model to attend to foreground objects while suppressing background. The optimization of the proposed model is formulated as a variational Bayesian inference problem, which is established by amortized inference networks.We conduct extensive experiments on three benchmarks, where our proposal obtains at least competitive and often better performance than state-of-the-art methods. We also provide comprehensive analyses and ablation studies to gain insight into the effectiveness of our method for few-shot semantic segmentation.

MetaKernel: Learning Variational Random Features with Limited Labels

Few-shot learning deals with the fundamental and challenging problem of learning from a few annotated samples, while being able to generalize well on new tasks. The crux of few-shot learning is to extract prior knowledge from related tasks to enable fast adaptation to a new task with a limited amount of data. In this paper, we propose meta-learning kernels with random Fourier features for few-shot learning, we call MetaKernel. Specifically, we propose learning variational random features in a data-driven manner to obtain task-specific kernels by leveraging the shared knowledge provided by related tasks in a meta-learning setting. We treat the random feature basis as the latent variable, which is estimated by variational inference. The shared knowledge from related tasks is incorporated into a context inference of the posterior, which we achieve via a long-short term memory module. To establish more expressive kernels, we deploy conditional normalizing flows based on coupling layers to achieve a richer posterior distribution over random Fourier bases. The resultant kernels are more informative and discriminative, which further improves the few-shot learning. To evaluate our method, we conduct extensive experiments on both few-shot image classification and regression tasks. A thorough ablation study demonstrates that the effectiveness of each introduced component in our method. The benchmark results on fourteen datasets demonstrate MetaKernel consistently delivers at least comparable and often better performance than state-of-the-art alternatives.

Direct Estimation of Spinal Cobb Angles by Structured Multi-Output Regression

The Cobb angle that quantitatively evaluates the spinal curvature plays an important role in the scoliosis diagnosis and treatment. Conventional measurement of these angles suffers from huge variability and low reliability due to intensive manual intervention. However, since there exist high ambiguity and variability around boundaries of vertebrae, it is challenging to obtain Cobb angles automatically. In this paper, we formulate the estimation of the Cobb angles from spinal X-rays as a multi-output regression task. We propose structured support vector regression (S^2VR) to jointly estimate Cobb angles and landmarks of the spine in X-rays in one single framework. The proposed S^2VR can faithfully handle the nonlinear relationship between input images and quantitative outputs, while explicitly capturing the intrinsic correlation of outputs. We introduce the manifold regularization to exploit the geometry of the output space. We propose learning the kernel in S2VR by kernel target alignment to enhance its discriminative ability. The proposed method is evaluated on the spinal X-rays dataset of 439 scoliosis subjects, which achieves the inspiring correlation coefficient of 92.76% with ground truth obtained manually by human experts and outperforms two baseline methods. Our method achieves the direct estimation of Cobb angles with high accuracy, which indicates its great potential in clinical use.

Learning to Learn Kernels with Variational Random Features

In this work, we introduce kernels with random Fourier features in the meta-learning framework to leverage their strong few-shot learning ability. We propose meta variational random features (MetaVRF) to learn adaptive kernels for the base-learner, which is developed in a latent variable model by treating the random feature basis as the latent variable. We formulate the optimization of MetaVRF as a variational inference problem by deriving an evidence lower bound under the meta-learning framework. To incorporate shared knowledge from related tasks, we propose a context inference of the posterior, which is established by an LSTM architecture. The LSTM-based inference network can effectively integrate the context information of previous tasks with task-specific information, generating informative and adaptive features. The learned MetaVRF can produce kernels of high representational power with a relatively low spectral sampling rate and also enables fast adaptation to new tasks. Experimental results on a variety of few-shot regression and classification tasks demonstrate that MetaVRF delivers much better, or at least competitive, performance compared to existing meta-learning alternatives.